The system of measuring the performance of a Web system using a workload generator can be modeled as a closed interactive system. In such a system, the throughput and the mean response time are related by the response time law. However, we find that a measured throughput and a corresponding measured mean response time can have significantly different accuracy. As a result, one metric may be more reliable than the other to identify the better of two given configurations of a Web system, which is an important problem that appears frequently in practice. Using simulation, we derive rules of thumb that characterize when throughput is more reliable than mean response time. Also, we explain these rules of thumb analytically. Specifically, we refine the response time law using the central limit theorem and formally define the asymptotic reliability of an estimator of a metric. Using these analytical frameworks, we provide insights into when and why one metric is more reliable than the other. © 2010 Elsevier B.V.